TL;DR
This paper proposes a narratologically grounded definition of "character" that emphasizes plot contribution, moving beyond simple animacy or entity recognition. They then demonstrate a straightforward supervised machine learning model that achieves promising results on this refined definition, particularly on a corpus of Russian folktales.
What Problem Does This Paper Address?
Understanding narratives is a key goal in natural language processing. A fundamental component of any narrative is its characters. However, existing computational approaches to character identification often lack a clear, theoretically grounded definition of what constitutes a character. They tend to either take character identification for granted, relying on heuristics for named entities or coreference chains, or use simplified definitions that don't capture the crucial distinction between a character and other animate entities in a story. Furthermore, many prior methods are overly complex, relying on extensive case bases or domain-specific ontologies. This paper aims to address this by proposing a more robust definition and demonstrating a practical machine learning approach.
Methods
The authors propose a two-stage approach to character identification:
- Animacy Detection: The first step is to identify animate entities within the narrative. They leverage an existing, state-of-the-art hybrid system that combines supervised machine learning with hand-built rules for this task. This is crucial because characters, by definition, must be animate.
- Character Classification: For the animate entities identified in the first stage, a supervised machine learning model is trained to classify them as characters. This model uses a straightforward Support Vector Machine (SVM) with a Radial Basis Function kernel.
The model explores various combinations of four key features:
- Coreference Chain Length (CL): The length of a coreference chain, hypothesizing that longer chains are more likely to represent characters.
- Semantic Subject (SS): Whether the head of a coreference chain acts as a semantic subject to a verb, indicating active participation.
- Named Entity (NE): Whether the head of a coreference chain is a PERSON named entity.
- WordNet (WN): Whether the head of a coreference chain is a descendant of "person" in WordNet, indicating a human-like entity.
The best-performing feature set, using all four features, was then applied to the datasets.
Datasets
The authors created preliminary annotated data for their experiments using two corpora:
- Extended ProppLearner Corpus: This corpus consists of 46 Russian folktales. It was originally annotated for major characters and referring expressions, coreference chains, and animacy, aligning well with the authors' proposed definition.
- InScript Corpus (Subset): A subset of 94 stories from the InScript corpus, focusing on narratives about taking a bath. This corpus provided annotated referring expressions and coreference chains. The authors manually annotated these chains for character status.
Key Insights and Ideas
- Character is more than just animate: The core insight is that a character is not merely an animate entity but one that actively contributes to the plot. This moves beyond simple animacy detection and focuses on functional roles within the narrative.
- Plot contribution is key: The authors hypothesize that features related to a referent's contribution to the plot will be critical for significantly improving character identification performance. This suggests a need for deeper semantic and narrative understanding.
- Simplicity can be effective: Despite the complexity of some prior approaches, the authors demonstrate that a straightforward supervised learning model with well-chosen features can achieve competitive results.
- Animacy is a necessary but not sufficient condition: The reliance on an animacy detector highlights that while characters must be animate, not all animate entities are characters. This distinction is crucial and a source of errors.
Key Results
The best performing model, utilizing all four features (CL, SS, NE, WN), achieved:
- An F₁ score of 0.81 on the ProppLearner corpus.
- An F₁ score of 0.99 on the InScript corpus.
The authors note that the InScript result is misleadingly high due to the simplicity of the stories (often first-person narratives with a single protagonist), making the task trivial. Error analysis on the ProppLearner data revealed that errors in animacy detection directly impact character identification, and that identifying characters with very few mentions remains a challenge.
Limitations and Future Work
The authors acknowledge several limitations and outline future directions:
- Animacy Model Dependency: The character identification model's performance is directly tied to the accuracy of the animacy detector. Improving animacy detection is a priority.
- Few Mentions: Identifying characters with very few mentions is difficult. The authors suggest incorporating features related to story events to address this.
- Distinguishing Characters from Animate Entities: Some non-character animate entities are misclassified as characters. Future work will focus on analyzing plot structure and developing features that are more specific to character identification than just animacy.
- Plot Contribution Features: The authors strongly believe that features encoding plot contribution will be critical for substantial improvements and plan to explore this in future work.
- Double Annotation: The current annotations were done by a single author; double annotation will be conducted in future work to ensure reliability.
Summary updated